---
title: 'HPGNN: Hierarchical & Higher-order GNN Models'
url: https://www.emergentmind.com/topics/hpgnn
type: topic
---

# HPGNN: Hierarchical & Higher-order GNN Models

HPGNN is an acronym used in the arXiv literature for more than one graph-neural architecture rather than a single canonical model. The two clearest exact usages are **Hierarchical Point Graph Neural Network**, introduced for large-scale outdoor LiDAR semantic segmentation [2206.02153], and **Higher-order Personalized PageRank Graph Neural Network**, introduced for heterophilic node classification [2507.16347]. In adjacent work, closely related labels such as **HP-GNN**, **HGNN**, **H-GNN**, and “**HPGNN-style**” are used for hardware-generation frameworks, hierarchical message-passing models, and hypergraph-based learning systems, which makes disambiguation technically important because the associated inductive biases are different [2112.11684][2107.07432][2508.01449].

## 1. Nomenclature and scope

The arXiv record suggests that “HPGNN” is best treated as a **polysemous acronym** rather than the name of one stable architecture family. One usage is tied to **hierarchical multiresolution processing of point clouds**, and another to **higher-order PPR-based propagation on heterophilic graphs**. A nearby but distinct label, **HP-GNN**, denotes a framework for automatically generating GNN training implementations on CPU-FPGA heterogeneous platforms rather than a new learning model.

| Label | Expansion | Primary setting |
|---|---|---|
| HPGNN | Hierarchical Point Graph Neural Network | Outdoor LiDAR semantic segmentation |
| HPGNN | Higher-order Personalized PageRank Graph Neural Network | Heterophilic node classification |
| HP-GNN | High-throughput GNN training implementation on CPU-FPGA Heterogeneous Platform | Hardware generation for GNN training |

This multiplicity matters because the corresponding methods solve different problems with different objects of computation. The point-cloud HPGNN operates on voxel-derived keypoint graphs across coarse and fine resolutions. The heterophily HPGNN operates on clique-complex liftings and higher-order PPR operators. The hyphenated HP-GNN, by contrast, is a systems framework whose inputs are GNN models, sampling algorithms, and platform parameters rather than graph data in the usual learning sense.

## 2. HPGNN as Hierarchical Point Graph Neural Network

In the point-cloud literature, HPGNN denotes **Hierarchical Point Graph Neural Network**, a purely GNN-based framework for large-scale outdoor LiDAR processing with single-scan semantic segmentation as the main demonstrated task. Its motivation is that standard point-level GNNs scale poorly on outdoor LiDAR and struggle to preserve fine structural detail when receptive fields are enlarged. The model therefore learns over multiple graph resolutions simultaneously, with a fine graph \(G_L\) preserving local structure and a coarse graph \(G_H\) capturing broader spatial context [2206.02153].

The graph construction is based on **cylindrical voxelization** in \((r,\theta,z)\) coordinates. Two voxel grids with different resolutions are built, each voxel produces a key point defined as the mean position of points inside the voxel, and these key points become nodes. At each level,
$$
G=(P,E),
$$
with points represented as
$$
p_u=(s_u,x_u),
$$
where \(x_u\) is spatial location and \(s_u\) is the feature vector. Edges are radius-based:
$$
E_i=\{(p_u,p_v)\mid \Delta_x<d_i\},\qquad \Delta_x=\|x_u-x_v\|_2.
$$

Within each graph, message passing follows the usual edge-message and node-update pattern:
$$
e^t_{uv}=f(s_u,s_v,\Delta_x),
$$
$$
s^{t+1}_u=g\left(\mathrm{Agg}\{e^t_{uv}\mid (u,v)\in E_i\}\right)+s^t_u.
$$
The intra-level aggregator is **Max aggregation**, chosen because it is described as less prone to oversmoothing than mean aggregation as neighborhoods grow. The residual term preserves prior features and stabilizes optimization.

The hierarchical mechanism is the central design feature. After \(T_1\) iterations on \(G_L\), fine features are transferred to \(G_H\) by a learned, attention-weighted downsampling operation. For a high-level point \(p\in P_H\) and contained fine points \(\{q_1,\dots,q_k\}\subset P_L\),
$$
e_{p,q_i}=h(s_{q_i},\Delta_x),
$$
$$
s^0_p=D\left(\Sigma(\alpha_i e_{p,q_i})\right).
$$
The coarse graph then performs \(T_2\) iterations of its own message passing, after which an upsampling layer \(U(\cdot)\) returns global features to \(G_L\). These are concatenated with fine-level features via a skip connection, and the fine graph is refined for \(T_3\) additional steps. The main model is a **2-level HPGNN with \((T_1,T_2,T_3)=(1,2,1)\)**.

For semantic segmentation, the loss combines weighted cross entropy, Lovasz-softmax loss, and \(L_2\) regularization:
$$
L_\mathrm{total}=\alpha L_\mathrm{wce}+\beta L_\mathrm{ls}+\gamma L_\mathrm{reg}.
$$
The weighted cross entropy is
$$
L_\mathrm{wce}=-\frac{1}{N}\sum_{i=1}^{N} w_c \ln(\hat{y}_i).
$$
The emphasis on class imbalance is important because outdoor LiDAR datasets contain small and vulnerable classes such as motorcyclist, bicyclist, and pole.

The reported results establish HPGNN mainly as a **graph-based baseline that closes much of the gap between early GNN methods and stronger segmentation pipelines**. On SemanticKITTI, the paper reports **SPGraph: 17.4 mIoU** and **HPGNN: 54.1 mIoU**, a gain of **36.7 mIoU**. On nuScenes-Lidarseg, it reports **63.8 mIoU**. At the same time, the model is explicitly not the best overall against specialized segmentation systems such as Cylinder3D, AF2S3Net, RPVNet, or SPVCNN++. The ablation study on SemanticKITTI sequence 08 is particularly diagnostic: \(G_L\) only with \((1,0,1)\) yields **43.3 mIoU**, \(G_H\) only with \((0,2,0)\) yields **41.0 mIoU**, and the two-level model with \((1,2,1)\) yields **51.6 mIoU**. This supports the paper’s claim that the benefit comes from combining resolutions rather than merely choosing one.

The main limitations are also explicit. Voxelization introduces discretization errors, especially near boundaries; dense graphs remain spatially inefficient at high resolution; and the method, as a GNN backbone, does not incorporate the task-specific fusion modules used by stronger point-, range-, or voxel-view segmentation systems.

## 3. HPGNN as Higher-order Personalized PageRank Graph Neural Network

In heterophily research, HPGNN denotes **Higher-order Personalized PageRank Graph Neural Network**, a model designed to mitigate the failure modes of standard neighbor aggregation when adjacent nodes frequently belong to different classes. The stated motivation is that existing heterophily methods, including APPNP, GPRGNN, and BernNet, primarily rely on pairwise propagation and overlook higher-order structures. HPGNN addresses this by combining **clique complex lifting**, **Higher-order Personalized PageRank (HiPPR)**, and **Higher-order Adaptive Spectral Convolution (HiASC)** [2507.16347].

The first step is to lift the original graph \(G=(V,E)\) into a simplicial complex \(\mathcal{K}\) by **clique complex lifting**:
$$
\mathcal{K}=\bigcup_{C_i\in C(G)} \sigma_i,
$$
where each clique \(C_i\) becomes a simplex \(\sigma_i\). This makes higher-order units such as triangles explicit, so propagation is no longer restricted to pairwise edges.

Standard PPR is recalled in fixed-point form as
$$
\pi_{\text{ppr}(i_s)}=\alpha i_s+(1-\alpha)\tilde{A}\pi_{\text{ppr}(i_s)},
$$
with matrix solution
$$
\Pi_{\text{PPR}}=\alpha\left(I_n-(1-\alpha)\tilde{A}\right)^{-1}.
$$
HPGNN generalizes this to higher-order simplex spaces:
$$
\Pi^{p}_{\text{HiPPR}}=\alpha \left(I^p-(1-\alpha)\tilde{A}^p\right)^{-1}.
$$
Here \(p\) indexes simplex order, so the propagation operator acts over higher-order topological objects rather than only nodes and edges.

Because direct inversion is expensive, the paper introduces **HiPwrPushSOR**, an approximation based on Push, Power Iteration, and Successive Over-Relaxation. The algorithm maintains an estimate \(\hat{\pi}\), a residual \(r\), and an active queue. It starts from
$$
r(s,s)=1,\qquad \hat{\pi}(s,\sigma)=0,
$$
and updates
$$
\hat{\pi}(s,\sigma)\leftarrow \hat{\pi}(s,\sigma)+\omega\alpha r(s,\sigma),
$$
$$
r(s,\tau)\leftarrow r(s,\tau)+\frac{\omega(1-\alpha)r(s,\sigma)}{d_\sigma},
$$
for outgoing higher-order neighbors \(\tau\in N_{\text{out}(\sigma)}\). The stated purpose is to make higher-order PPR computation practical while preserving the denoising behavior associated with restart-based propagation.

The convolutional layer is formulated through repeated application of normalized HiPPR operators:
$$
H^{0}=f(X),\qquad H^{k}= \tilde{\Pi}^p_{\text{HiPPR}} H^{(k-1)},
$$
with
$$
\tilde{\Pi}^p_{\text{HiPPR}}=D^{-1/2}\Pi^p_{\text{HiPPR}}D^{-1/2}.
$$
Order-specific outputs are combined as
$$
Z^p=\sum_{k=0}^{K}\beta_{p,k}H^{p,k},
$$
and the final representation is
$$
Y=\rho\left(\big\|_{p=1}^{P} Z^p\right).
$$
The linearized form given in the paper is
$$
Y=\big\|_{p=1}^{P}\left(\sum_{k=0}^{K}\beta_{p,k}\tilde{\Pi}^p_{\text{HiPPR}}X\Theta_p\right)W.
$$

The empirical setting covers seven real-world datasets. On the heterophilic benchmarks, the reported homophily ratios are low: **Cornell 0.127**, **Actor 0.219**, **Texas 0.087**, and **Wisconsin 0.192**. HPGNN reports the following accuracies: **Cora 89.96**, **Citeseer 79.89**, **Photo 95.19**, **Cornell 88.30**, **Actor 40.89**, **Texas 91.48**, and **Wisconsin 94.75**. The paper states that the model achieves better performance than five out of seven state-of-the-art methods on heterophilic graphs while remaining competitive on homophilic graphs.

Ablation results isolate the roles of higher-order structure and PPR. Removing higher-order structure, or removing both higher-order structure and PPR, lowers performance, especially on heterophilic datasets such as Actor and Wisconsin. Parameter studies reported in the paper adopt **\(p=2\)** as the default order, **\(\alpha=0.15\)** as a practical teleport choice, and **\(k=10\)** as a good tradeoff for the convolution depth. The conceptual claim is that higher-order context and personalized restart are complementary: higher-order topology adds expressiveness, while PPR controls noisy over-mixing.

## 4. Relation to adjacent hierarchical and hypergraph models

The broader literature around HPGNN spans several nearby but non-identical design lines. **Hierarchical Graph Net (HGNet)** studies hierarchical message passing for long-range interactions and proves that, for any two connected nodes, there exist message-passing paths of at most logarithmic length with respect to graph size, while the total hierarchy remains asymptotically equivalent in size to the input graph under mild assumptions [2107.07432]. This is conceptually close to the point-cloud HPGNN, which also uses a fine-to-coarse-to-fine communication pattern, but HGNet is not a point-cloud model and is motivated by long-range interactions in generic graphs.

In hypergraph learning, **Tensorized Hypergraph Neural Networks (THNN)** position themselves explicitly against mainstream **HGNN / HPGNN-style models** that collapse a hypergraph into a graph-like first-order surrogate. THNN instead uses an adjacency tensor and high-order outer-product message passing,
$$
x_i^{(l+1)} = \sigma\!\left( \sum_{(j_1,\ldots,j_{N-1})\in \mathcal{N}_i} \left(x_{j_1}^{(l)} \circ \cdots \circ x_{j_{N-1}}^{(l)}\right) \times_{(2,3,\ldots,N)} W \right),
$$
with partially symmetric CP decomposition to reduce complexity from exponential to linear in rank and tensor order [2306.02560]. The comparison is important because it shows that “higher-order” can mean either multiresolution hierarchy, as in point-cloud HPGNN, or explicit multilinear hyperedge modeling, as in THNN.

A separate line, **HWL-HIN**, makes injectivity the organizing principle. It performs dual-stage node-to-hyperedge and hyperedge-to-node updates with GIN-style sums and proves expressive power exactly equivalent to the Hypergraph Weisfeiler-Lehman test [2512.22014]. Its core layer updates,
$$
h_e^{(l+1)}=\text{MLP}_e^{(l)}\left((1+\epsilon_e^{(l)})h_e^{(l)}+\sum_{v_i\in\mathcal{N}_v(e)}h_{v_i}^{(l)}\right),
$$
$$
h_v^{(l+1)}=\text{MLP}_v^{(l)}\left((1+\epsilon_v^{(l)})h_v^{(l)}+\sum_{e_i\in\mathcal{N}_e(v)}h_{e_i}^{(l+1)}\right),
$$
occupy yet another point in the nearby design space. A plausible implication is that “HPGNN” in current usage often marks a **problem-specific higher-order or hierarchical inductive bias**, not a uniform architectural template.

## 5. Common confusions and nearby acronyms

The term should not be conflated with several adjacent names.

**HP-GNN** is a hardware framework, not a learning architecture. It automatically generates sampling-based mini-batch GNN training implementations for CPU-FPGA platforms, using data-layout optimizations, HLS hardware templates, a design space exploration engine, and high-level Python APIs. It reports **55.67× average speedup over CPU-only**, **2.17× average speedup over CPU-GPU**, and up to **4.45× speedup over GraphACT** [2112.11684].

**HGAttack** is not a new heterogeneous GNN at all, but a **targeted gray-box evasion attack** against heterogeneous graph neural networks. Its contribution lies in a meta-path-based surrogate model and a semantics-aware gradient mechanism for perturbation generation, and it is explicitly framed as an attack benchmark rather than an HGNN architecture [2401.09945].

**VisitHGNN** is a heterogeneous, relation-specific GNN for predicting origin distributions of POI visits over candidate CBGs. It uses CBG adjacency, POI-POI geo/time/brand relations, cross-type fusion, and a masked KL objective, and reports **KL 0.287**, **MAE 0.0082**, **Top-1 0.853**, and **\(R^2=0.892\)** at \(K=50\) [2510.02702]. Despite the “HGNN” suffix, it is neither of the two architectures denoted by the exact acronym HPGNN.

Similarly, **H-GNN** in recent collider phenomenology denotes a **Hyper-Graph Neural Network** for \(t\bar t t\bar t\) classification, where nodes are jets, leptons, and \(p_T^{\rm miss}\), and hyperedges encode pairwise, three-body, and four-body correlations. It attains **AUC \(=0.951\)** and **\(Z=9.11\)** at \(140~\mathrm{fb}^{-1}\) [2605.18382]. This is a genuine hypergraph model, but not the same object as either point-cloud HPGNN or PPR-based HPGNN.

A recurring misconception is therefore to read “HPGNN” as a generic synonym for “hypergraph neural network.” The literature provided here does not support that equivalence. Instead, it supports a narrower conclusion: the acronym is reused across distinct subfields, while hypergraph models more often appear under labels such as **HNN**, **HWL-HIN**, **THNN**, or **H-GNN**.

## 6. Significance in the broader GNN landscape

The broader HGNN survey literature organizes hypergraph neural networks into **hypergraph convolutional networks**, **hypergraph attention networks**, **hypergraph autoencoders**, **hypergraph recurrent networks**, and **deep hypergraph generative models**, emphasizing higher-order relational learning, scalability, hypergraph construction, robustness, and temporal modeling as common themes [2503.07959]. Against that taxonomy, the two exact HPGNN usages occupy different positions: the LiDAR model is fundamentally **hierarchical and multiresolution**, whereas the heterophily model is **higher-order, PPR-based, and simplicial**.

Recent application work also shows why the naming drift persists. In collider analysis, graph-based event representations are motivated by the claim that meaningful topology—pairing relations among reconstructed objects, angular separations, and hierarchical decay structure—is more naturally encoded by geometry and connectivity than by a flat tabulation of engineered features. A graph neural network built on such a representation improves the expected 95% CL upper limit on the double-Higgs production cross-section by **28%** relative to an XGBoost baseline in \(HH\to b\bar b\gamma\gamma\) [2508.01449]. This suggests that “HPGNN-style” language can function as a shorthand for **architectures that explicitly exploit higher-order, hierarchical, or geometric structure**, even when the exact model name differs.

The main encyclopedia-level conclusion is therefore terminological as much as architectural. In current arXiv usage, **HPGNN does not denote one settled model class**. It denotes, most prominently, either a **Hierarchical Point Graph Neural Network** for outdoor LiDAR segmentation or a **Higher-order Personalized PageRank Graph Neural Network** for heterophily mitigation. Around these two exact meanings lies a larger neighborhood of hierarchical, higher-order, and hypergraph methods whose shared premise is that pairwise, flat message passing is often insufficient, but whose implementations and theoretical commitments differ substantially.

Source: https://www.emergentmind.com/topics/hpgnn